Lightweight Image Super-Resolution Preprocessor for Jpeg Compression

This paper proposes a lightweight image super-resolution (SR) pre-processing method for JPEG compression. Our method reduces computational costs for super-resolution and ensures consistent image quality by avoiding unknown super-resolution algorithms on the user side, which may cause inconsistencies in the quality of the delivered images. Specifically, the proposed method contains an SR model with only 2.8K parameters, followed by the JPEG encoder. For the lightweight SR model, a Separable Structural Reparameterization (SSR), which consists of a Cross-Scale Directional Reparam (CDR) module that improves spatial learning and a Channel Reparam Integration (CRI) module that enhances feature interaction, is proposed to improve the capability of the model for information enhancement. Furthermore, we introduce a JPEG compression-aware training strategy to make the super-resolution process aware of the subsequent JPEG compression. This strategy uses a differentiable JPEG codec for gradient backpropagation during training and switches to the standard JPEG codec for deployment. Extensive experiments have been conducted to validate the effectiveness of our method.